Inside the machine room

monefy is a team in Brașov, Romania that runs the full back office of several US healthcare operations: billing, CRM, call centre, marketing, and the software that holds it all together. Our working day starts at three in the afternoon so it can end at midnight, matching the working day of the operations we run.

Over the last two years we built ourselves an operating system for that work. Automations that press the buttons humans should not have to. Data pipelines that re-derive each other’s arithmetic before anyone hits send. Internal agents with names, job descriptions, and the standing authority to tell us no, including the time one of them blocked a launch of ours, on launch day, with citations.

We have not written about any of it. That is what this is.

Why now

Mostly because the gap became embarrassing. Our website said we were a competent back-office provider, which is true and roughly as interesting as a dial tone. Meanwhile the actual work involves a team teaching itself a foreign country’s insurance rules from scratch, a ten-volume internal curriculum on American culture, and a billing robot that spent its first month pressing exactly one button under supervision.

The second reason is hiring. We are competing for engineers in a city with a lot of engineering jobs, and “we do BPO” is a sentence that closes a conversation. What we actually do is stranger and more interesting than that, and the only honest way to prove it is to show the work.

What the machine actually is

Not one system. A stack of habits with tooling underneath, most of it unremarkable in isolation:

• Nothing is trusted once. Any number that reaches a client or a decision gets derived a second time, by a different path, before it counts. It is exactly as tedious as it sounds. It has also caught things that would have been very expensive to find later.
• Automation earns permissions. Every robot starts deliberately boring, one button, small batch, hand-checkable. It widens only after a month of being right.
• Plans get attacked before they get built. We point a group of agents at a proposal with instructions to tear it apart from several angles at once. Most plans come back smaller.
• Policy ships like software. Written down, versioned, reviewed, dated. If a rule exists only in someone’s head, it does not exist.

The first version of every robot we ship is deliberately boring. Half the value, a hundredth of the risk, then it earns the next permission.

One concrete example

Insurance remittances look like chaos. Thousands of lines a month, hundreds of adjustment codes, every payer with its own dialect. The instinct is to hire more people to read them.

We sorted a month of them instead. 9,734 remittance lines resolved into eleven workflows, once we had mapped the 177 recurring reason-code patterns underneath them.

The map lives in a file our billing team maintains, not in code somebody has to be paged about. And here is the part that actually matters: anything the map does not recognise goes straight to a person, automatically. Not to a log file, not to a retry queue, not to a dashboard nobody opens. To a human, by default, immediately.

That last sentence is the whole design. It is easy to build something that handles 95% of cases and quietly mishandles the rest. The unglamorous work is making the 5% loud.

What we will write about

Things with an artifact at the centre, a tool, a ritual, an incident, a decision, and an honest account of why it exists. Some of it will be engineering. Some will be about teaching a team in Romania how small talk works five thousand miles away, which turns out to be a genuine operational problem when your staff answer phones on another continent.

What we will not write: client names, anything patient-adjacent, revenue figures, or the parts of our commercial playbook we would rather competitors kept guessing at. We operate other people’s businesses. Most of what is interesting is not ours to publish, and the machine is transferable in a way the specifics are not.

If you run an operation that has grown past what memory and manual steps can hold, or if you would rather build the machine than be described by it, those are two different conversations and we are up for both.